Psychology 209 – 2017

Homework #2 – due Jan 24, 2017 (version 2, Wed, Jan 18, 10:27 am)

The purpose of this homework is two-fold:

1. To get you set up with the PDPyFlow software and some of the tools you will be able to use in this and later homeworks. The key elements here are

The PDPyFlow software background, philosophy and current status

The management of your copy of the PDPyFlowsoftware

The minimal PDPyFlow user interface and graphical tools we have provided

2. To allow you to explore several key concepts relied on in our conceptual framework for understanding the neural basis of perception. The key concepts here are

Generative model of an environmentand its instantiation in a neural network

How a neural network samples from the probability distribution of the generative model

How this approach allows us to:

(i) exploit conjunctions of constraints rather than simply their additive combination and
(ii) exhibit sensitivity to the implicit statistics of the training set

The first two pages below are preliminaries:

-Background on PDPyFlow

-Setting up to run your simulations

This is followed by the Homework handout itself, which builds on McClelland (2013), and then takes you into the actual simulation and questions you are to answer there

The last page is a User’s guide for the software tools used for this assignment

PDPyFlow: Background, Philosophy and Current Status

In the late 1980’s I developed the PDP software, implementing models in the PDP volumes. This software was written in C, ran on the simple PC’s available at that time, and used a 24 x 80 character window for both the user interface and for visualization. The software was re-implemented in Matlab in the late 1990’s. Today, we have moved to Python and Tensorflow for two reasons:First, these tools are open source resources while MATLAB is a proprietary product. Second, Tensorflow, which is built on top of Python, is the strongest tool for neural network research, allowing us to use GPU’s without thinking about them. Our approach is intended to allow you to begin to work with this cutting edge computational toolkit even if you have no prior experience or background with neural networks.

PDPyFlow is currently a work in progress. Alex Ten () a friendly and capable young man from Kazakhstan is doing the implementation. If you encounter bugs, contact Alex.

Managing your copy of PDPyFlow.Your account on the PDP lab computer will come pre-configured with a copy of the modules of the PDPyFlow software. Currently there are three application modules and some support software. The application modules are the MIA module, the FFBP module, and the RNN module. For the first homework, we will be working with the MIA module.

You will have your own personal copy of the software in your own directory on the lab’s computer. This will allow you to make modifications on top of the existing software and to save your changed versions. An advantage of this is that you have total control – and, if you are a great developer, you could even contribute to the enhancement of the software by creating or extending tools. Given that the software is still being developed, you may need to update your copy before you begin a homework assignment. In general, you should make sure you have the latest copy of everything by executing the ‘git pull’ command once you are inside the PDPyFlow directory (currently just called PDP). If you find yourself wishing to make extensions to the code, please consult with Alex.

Minimal user interface with helpful graphical tools.Many software systems rely on a GUI. The advantage of this is that you don’t have to remember long lists of parameter and command conventions. The disadvantage is that this approach invariably hides the inner structure of your software from you. With PDPyFlow we have opted to try to make everything accessible to you at the cost of requiring you to type commands at the command prompt. This is facilitated if you use a text editor to inspect the code so you can see how things are named. That way you can even make modifications to defaults used in the code so you will only need to type a minimum of commands at the prompt at run time. For the first homework, we provide a simple user’s guide at the end of this document, so editing files will not be necessary.

As a compromise to allow you to run exercises without having to learn arbitrary things, we have set things up to make basic use simple. For the first homework, you will need only a few simple commands. Extension then is open to you – though for that, you’ll have to do some digging for yourself. Detailed documentation of some things is unfortunately not yet available.

On the other hand, because there are particular quantitative simulation results and values that we want you to be able to track,we have created visualization tools that make the values we want you to understand accessible in a window that pops up and allows you to navigate around in your results. We’ll describe the version of this window that we will be working withbelow.

Setting up to Run your own Simulations

Psychology 209 uses a small server with GPUs owned by the PDP lab. Use these instructions to access the lab computer to do the exercises for the second and subsequent homework. Contact Steven () if you encounter difficulties.

Logging onto the server:

  1. Must be connected to the Stanford network

-- if working from off campus, or from a laptop, you may need to use the VPN

  1. Mac users:
  2. Open the terminal
  3. ssh -Y
  4. Linux users:
  5. ssh -X
  6. Windows users
  7. onetime setup:
  8. Download and install Putty
  9. Download and install Xming
  10. Launch Xming
  11. Launch and Configure Putty
  12. Type into the hostname field
  13. In the “ssh” tab, find the X11 field and select the remote X11 authentication protocol with MIT magic cookie
  14. Type a name for this configuration (e.g psych209), in Saved Sessions and save for future use
  15. Load saved configuration and press open
  16. Regular use:
  17. Launch Xming
  18. Lauch Putty, Load saved configuration, press open
  19. When it asks for the password, it is “password”
  20. Immediately change your password with something more secure.
    Enter passwd at the command prompt

You will be prompted for your existing password, then for your new password

  1. The PDP folder contains your copy of the code needed for the homework exercises
  2. Use the “cd” command (change directory) to enter this directory: cd PDP
  3. To ensure you have the most recent updates, use the git pull command from inside the PDP directory before you start your using the software
  4. This system has Tensorflow installed, and has two GPUs that Tensorflow uses. Please don’t use Tensorflow until we’ve gone over it in class, as by default a single user will use an entire GPU (not good unless authorized!)

Homework Background:
Generative model of an environment and its relation to the state of a neural network

We assume here that you have finished reading the McClelland (2013) paper that describes the relationship between neural network models of perception and Bayesian inference. Here we briefly summarize a couple of the key points and organize them to allow you to focus on the right concepts for the homework.

We envision a world of visual stimuli that are generated from a vocabulary of 36 different words. The words are shown in the table below, and come from an experiment by Movellan and McClelland (2001), and some of the materials were used as examples in McClelland (2013). Note that three of the words from M&M (01) were not used (the ones with a line through them), and the word WAR occurs in the table twicebut is only considered to be a single word in the model. Although the words have different frequencies in the English language (and these are shown in the table), we do not incorporate this variable into the generative model, for simplicity.

In the generative model for this situation, stimuli are created by first selecting a word from the vocabulary (each word is selected with equal probability), then generating a set of three letters conditioned on the word and finally generating a set of features conditioned on the generated letters. For example, on a given trial we might select the word AGE,then select the three letters H, G, and E, and then select features based on the letters. A parameter of the model, called ‘OLgivenW’ determines the probability under the model that we select the correct letter in each position given the word that was chosen under the generative model. Let’s understand this since it can be confusing. Let’s say OLgivenW is 10. What this means is that the odds are ten to one that we will select the correct letter rather than a given incorrect letter. For example, we are 10 times more likely to select A in the first position than we are to select H, according to this generative model.

With this value of OLgivenW, we will actually choose one of the incorrect letters more often than the correct letter, since there are 25 incorrect letters.
The probability that we will chose the correct letter is OLgivenW/(OLgivenW + 25). So if OLgivenW is 10, the probability that we will select the correct letter is only 10/35, or about .286. You should be able to see that the probability that we would choose H in position 0 is .0286.

We can already calculate the probability under the generative model that we would have selected the word AGE and then generated the letters H, G, and E. We assume conditional independence, that is, that the letters are chosen conditional on the word AGE but independently for each position, so their choice is conditionally independent given the word AGE. The calculated probability is 1/36 (the probability of selecting AGE) times .0286 (the probability of generating h in position 0 given age), times .286 (the probability of generating g in position 1 given age) times .286 (the probability of generating e in position 2 given age. This turns out to be a pretty small number: 6.4788e-05, or .000064788. What is the probability that we would have selected AGE and then generated all three letters correctly? What is the probability that we would have selected ‘age’ and then generated the letters H, X, and K? You can determine these additional probabilities using OLgivenW and using the assumption of conditional independence. (Answers are 6.4788e-04 and 6.4788e-07. Be sure you see why).

Next, we assume that the features that a participant sees are also probabilistic, with the features in each position chosen independently conditional on the letter. Another parameter governs this: it is called OFgivenL, and by default we also chose a value of 10 for this parameter. That means that under the generative model, we are 10 times more likely to choose the value of a feature that is correct for a given letter than the value that is incorrect for that letter. For example, if the letter we have generated for position 0 is H, then we are 10 times more likely to specify that the vertical segment at the lower left of the first letter position is present (since this feature is present in the letter H) than to specify that it is absent. We are also 10 times more likely to specify that a horizontal feature at the top is absent than to specify that it is present.

With these given assumptions, let’s consider the probability that all of the features of a given letter will be generated correctly given the letter under the generative model. First, we note that the probability that a given feature is generated correctly is OFgivenL/(OFgivenL + 1), since there’s only one incorrect value for each feature. If OFgivenL is 10, this number is 10/11, or .9091.There are 14 of these features,and each is generated conditionally independently, so that the probability that all the features would be generated correctly is only moderate– it is (10/11)14, or .2633. What is the probability that all the values of the features will be correct, except for the bar at the top? It is (10/11)13*(1/11), or .02633 – the ratio of the probabilities of these cases is 10 to 1.

Under all of the assumptions of our generative model, and the given values of OLgivenW and OFgivenL, the probability that all the letters and all the features of our selected world would be generated correctly is actually quite low. But, on the other hand, correct letters are far more likely than incorrect, and the same is also true of features. Thus if we believe this sort of generative model underlies our experiences of the world, this will lead us to use context to interpret the inputs that we receive.

Our neural network. We now consider the neural network that instantiates the above generative model. In this neural network, we have a set of 36 word units, and we can set the bias weight for each equal to the log(1/36), since each word is equally probable. In addition, we have weights in the network from words to letters, and we can set these weights as follows: The weights from the word level to the letter level could be set equal to log(OLgivenW/(OLgivenW+25)) for the correct letter in each position and to log(1/(OLgivenW+25)) for all the incorrect letters. However, we can simplify here because when we compute we will be using the softmax function. Under the softmax function, the only things that matter are the ratios of corresponding probabilities, rather than their absolute values. For example, we can ignore the bias weights at the word level, because their ratios are all equal to 1, and log(1) is 0. Similarly, we can set the weight to the correct Letter from a given word to log(OLgivenW) and the weight to each incorrect letter from a given word to be equal to log(1) – ie we can let the incorrect weights all simply be equal to 0. Also, at the letter level, the weights between the letters and their correct feature values are given by log(OFgivenL), and the weights between letters and their incorrect feature values are again log(1), and thus are just left equal to 0. Our network then simplifies to one in which there are positive weights between mutually consistent units across levels and there is competition via the softmax function between units within a level. While these weights are defined by the generative model (i.e., top down), they are used in both directions in our computations. Between layer influences are exclusively excitatory (as in the brain, where connections between different regions in the brain are only excitatory), while inhibition is needed only to resolve competition among neurons within layer (again, in the brain, inhibition is local within a small region of cortex, similar to the way it is in the model).

In summary, the neural network model we will use to instantiate the generative model described above consists of a set of 36 word units – one for each of the words in the above table – and three pools of 26 letter units. Each pool of letter units has connections to its own set of 14 sets of mini-pools of feature level units, although in the code these are just 28 values of an input vector, since they are specified as inputs by the modeler. The letter pools are indexed 0, 1, and 2, because in Python and Tensorflow we count from 0. There are no bias terms in the network, but there are weights that link word units to the letters that are actually in these words, and weights that link letter units to the values of features that correctly characterize a letter (thus there’s a weight from A to ‘crossbar at the top present’ and a weight from H to ‘crossbar at the top absent’) There are only two free parameters that determine the values of the weights: Weights between words and their letters are equal to log(OLgivenW) and weights between letters and their features are equal to log(OFgivenL).

Processing in the network. Our neural network is thought of as sampling from the probability distribution of possible states than might have produced a given pattern of experienced features. For concreteness, we continue with the example of the input ‘hge’. What might have generated this under our generative model? Well, there are two particularly likely possibilities. One is that the word was AGE, but the first letter was mis-generated as an H rather than an A, and then the correct features of H were all generated. The other is that the word was AGE, all three of the correct letters were generated, and then the feature at the top of position 0 was incorrectly generated. If OLgivenW and OFgivenL are equal to each other, these two possibilities are actually equally likely. Either way, we should end up thinking that the underlying word was likely to be AGE, although the letter that gave rise to the features might have been either A or H.

Processing takes place as follows. We first set the features to the values specified in the input, and we make sure no units are already active at either the letter or the word level. Then, processing takes place over a series of timesteps (default:20). In each time step, we first update the units at the letter level in each position. We think of this as a parallel computation: For each letter in each position we calculate its net input, which is the number of features of each letter that are specified correctly in the input, scaled by log(OFgivenL). On timestep 0, there is no word level activity to affect this computation, so that’s all there is to it. In our example, the number of such features is 13 for A in the first position and 14 for H in the first position. The resulting net inputs are put through the softmax function and one letter is selected to be active with the probability specified by the output of the softmax. In the model, this is a discrete event. However, we actually run a large batchof independent samples (default: 1000) of such computations and we report the proportion of times each letter is selected across the batch. This is a random process, so that the value should only approximate the true value (If the true probability is p, and B is the batchsize, the sampled proportion will fall within the range

95% of the time. For B = 1000,the range is about ± .03 if p is around .5,± .02 if p is around .1 or .9, ± .01 if p is near .98 or .02, and less than ±.01 if p less than .01 or greater than .99).